def initialize(self, pc): """Set up the problem context. Take the original mixed problem and reformulate the problem as a hybridized mixed system. A KSP is created for the Lagrange multiplier system. """ from firedrake import (FunctionSpace, Function, Constant, TrialFunction, TrialFunctions, TestFunction, DirichletBC) from firedrake.assemble import (allocate_matrix, create_assembly_callable) from firedrake.formmanipulation import split_form from ufl.algorithms.replace import replace # Extract the problem context prefix = pc.getOptionsPrefix() + "hybridization_" _, P = pc.getOperators() self.ctx = P.getPythonContext() if not isinstance(self.ctx, ImplicitMatrixContext): raise ValueError( "The python context must be an ImplicitMatrixContext") test, trial = self.ctx.a.arguments() V = test.function_space() mesh = V.mesh() if len(V) != 2: raise ValueError("Expecting two function spaces.") if all(Vi.ufl_element().value_shape() for Vi in V): raise ValueError("Expecting an H(div) x L2 pair of spaces.") # Automagically determine which spaces are vector and scalar for i, Vi in enumerate(V): if Vi.ufl_element().sobolev_space().name == "HDiv": self.vidx = i else: assert Vi.ufl_element().sobolev_space().name == "L2" self.pidx = i # Create the space of approximate traces. W = V[self.vidx] if W.ufl_element().family() == "Brezzi-Douglas-Marini": tdegree = W.ufl_element().degree() else: try: # If we have a tensor product element h_deg, v_deg = W.ufl_element().degree() tdegree = (h_deg - 1, v_deg - 1) except TypeError: tdegree = W.ufl_element().degree() - 1 TraceSpace = FunctionSpace(mesh, "HDiv Trace", tdegree) # Break the function spaces and define fully discontinuous spaces broken_elements = ufl.MixedElement( [ufl.BrokenElement(Vi.ufl_element()) for Vi in V]) V_d = FunctionSpace(mesh, broken_elements) # Set up the functions for the original, hybridized # and schur complement systems self.broken_solution = Function(V_d) self.broken_residual = Function(V_d) self.trace_solution = Function(TraceSpace) self.unbroken_solution = Function(V) self.unbroken_residual = Function(V) shapes = (V[self.vidx].finat_element.space_dimension(), np.prod(V[self.vidx].shape)) domain = "{[i,j]: 0 <= i < %d and 0 <= j < %d}" % shapes instructions = """ for i, j w[i,j] = w[i,j] + 1 end """ self.weight = Function(V[self.vidx]) par_loop((domain, instructions), ufl.dx, {"w": (self.weight, INC)}, is_loopy_kernel=True) instructions = """ for i, j vec_out[i,j] = vec_out[i,j] + vec_in[i,j]/w[i,j] end """ self.average_kernel = (domain, instructions) # Create the symbolic Schur-reduction: # Original mixed operator replaced with "broken" # arguments arg_map = {test: TestFunction(V_d), trial: TrialFunction(V_d)} Atilde = Tensor(replace(self.ctx.a, arg_map)) gammar = TestFunction(TraceSpace) n = ufl.FacetNormal(mesh) sigma = TrialFunctions(V_d)[self.vidx] if mesh.cell_set._extruded: Kform = (gammar('+') * ufl.jump(sigma, n=n) * ufl.dS_h + gammar('+') * ufl.jump(sigma, n=n) * ufl.dS_v) else: Kform = (gammar('+') * ufl.jump(sigma, n=n) * ufl.dS) # Here we deal with boundaries. If there are Neumann # conditions (which should be enforced strongly for # H(div)xL^2) then we need to add jump terms on the exterior # facets. If there are Dirichlet conditions (which should be # enforced weakly) then we need to zero out the trace # variables there as they are not active (otherwise the hybrid # problem is not well-posed). # If boundary conditions are contained in the ImplicitMatrixContext: if self.ctx.row_bcs: # Find all the subdomains with neumann BCS # These are Dirichlet BCs on the vidx space neumann_subdomains = set() for bc in self.ctx.row_bcs: if bc.function_space().index == self.pidx: raise NotImplementedError( "Dirichlet conditions for scalar variable not supported. Use a weak bc" ) if bc.function_space().index != self.vidx: raise NotImplementedError( "Dirichlet bc set on unsupported space.") # append the set of sub domains subdom = bc.sub_domain if isinstance(subdom, str): neumann_subdomains |= set([subdom]) else: neumann_subdomains |= set( as_tuple(subdom, numbers.Integral)) # separate out the top and bottom bcs extruded_neumann_subdomains = neumann_subdomains & { "top", "bottom" } neumann_subdomains = neumann_subdomains - extruded_neumann_subdomains integrand = gammar * ufl.dot(sigma, n) measures = [] trace_subdomains = [] if mesh.cell_set._extruded: ds = ufl.ds_v for subdomain in sorted(extruded_neumann_subdomains): measures.append({ "top": ufl.ds_t, "bottom": ufl.ds_b }[subdomain]) trace_subdomains.extend( sorted({"top", "bottom"} - extruded_neumann_subdomains)) else: ds = ufl.ds if "on_boundary" in neumann_subdomains: measures.append(ds) else: measures.extend((ds(sd) for sd in sorted(neumann_subdomains))) markers = [int(x) for x in mesh.exterior_facets.unique_markers] dirichlet_subdomains = set(markers) - neumann_subdomains trace_subdomains.extend(sorted(dirichlet_subdomains)) for measure in measures: Kform += integrand * measure trace_bcs = [ DirichletBC(TraceSpace, Constant(0.0), subdomain) for subdomain in trace_subdomains ] else: # No bcs were provided, we assume weak Dirichlet conditions. # We zero out the contribution of the trace variables on # the exterior boundary. Extruded cells will have both # horizontal and vertical facets trace_subdomains = ["on_boundary"] if mesh.cell_set._extruded: trace_subdomains.extend(["bottom", "top"]) trace_bcs = [ DirichletBC(TraceSpace, Constant(0.0), subdomain) for subdomain in trace_subdomains ] # Make a SLATE tensor from Kform K = Tensor(Kform) # Assemble the Schur complement operator and right-hand side self.schur_rhs = Function(TraceSpace) self._assemble_Srhs = create_assembly_callable( K * Atilde.inv * AssembledVector(self.broken_residual), tensor=self.schur_rhs, form_compiler_parameters=self.ctx.fc_params) mat_type = PETSc.Options().getString(prefix + "mat_type", "aij") schur_comp = K * Atilde.inv * K.T self.S = allocate_matrix(schur_comp, bcs=trace_bcs, form_compiler_parameters=self.ctx.fc_params, mat_type=mat_type, options_prefix=prefix) self._assemble_S = create_assembly_callable( schur_comp, tensor=self.S, bcs=trace_bcs, form_compiler_parameters=self.ctx.fc_params, mat_type=mat_type) with timed_region("HybridOperatorAssembly"): self._assemble_S() Smat = self.S.petscmat nullspace = self.ctx.appctx.get("trace_nullspace", None) if nullspace is not None: nsp = nullspace(TraceSpace) Smat.setNullSpace(nsp.nullspace(comm=pc.comm)) # Set up the KSP for the system of Lagrange multipliers trace_ksp = PETSc.KSP().create(comm=pc.comm) trace_ksp.setOptionsPrefix(prefix) trace_ksp.setOperators(Smat) trace_ksp.setUp() trace_ksp.setFromOptions() self.trace_ksp = trace_ksp split_mixed_op = dict(split_form(Atilde.form)) split_trace_op = dict(split_form(K.form)) # Generate reconstruction calls self._reconstruction_calls(split_mixed_op, split_trace_op)
def initialize(self, pc): """Set up the problem context. Take the original mixed problem and reformulate the problem as a hybridized mixed system. A KSP is created for the Lagrange multiplier system. """ from firedrake import (FunctionSpace, Function, Constant, TrialFunction, TrialFunctions, TestFunction, DirichletBC, assemble) from firedrake.assemble import (allocate_matrix, create_assembly_callable) from firedrake.formmanipulation import split_form from ufl.algorithms.replace import replace # Extract the problem context prefix = pc.getOptionsPrefix() + "hybridization_" _, P = pc.getOperators() self.cxt = P.getPythonContext() if not isinstance(self.cxt, ImplicitMatrixContext): raise ValueError("The python context must be an ImplicitMatrixContext") test, trial = self.cxt.a.arguments() V = test.function_space() mesh = V.mesh() if len(V) != 2: raise ValueError("Expecting two function spaces.") if all(Vi.ufl_element().value_shape() for Vi in V): raise ValueError("Expecting an H(div) x L2 pair of spaces.") # Automagically determine which spaces are vector and scalar for i, Vi in enumerate(V): if Vi.ufl_element().sobolev_space().name == "HDiv": self.vidx = i else: assert Vi.ufl_element().sobolev_space().name == "L2" self.pidx = i # Create the space of approximate traces. W = V[self.vidx] if W.ufl_element().family() == "Brezzi-Douglas-Marini": tdegree = W.ufl_element().degree() else: try: # If we have a tensor product element h_deg, v_deg = W.ufl_element().degree() tdegree = (h_deg - 1, v_deg - 1) except TypeError: tdegree = W.ufl_element().degree() - 1 TraceSpace = FunctionSpace(mesh, "HDiv Trace", tdegree) # Break the function spaces and define fully discontinuous spaces broken_elements = ufl.MixedElement([ufl.BrokenElement(Vi.ufl_element()) for Vi in V]) V_d = FunctionSpace(mesh, broken_elements) # Set up the functions for the original, hybridized # and schur complement systems self.broken_solution = Function(V_d) self.broken_residual = Function(V_d) self.trace_solution = Function(TraceSpace) self.unbroken_solution = Function(V) self.unbroken_residual = Function(V) # Set up the KSP for the hdiv residual projection hdiv_mass_ksp = PETSc.KSP().create(comm=pc.comm) hdiv_mass_ksp.setOptionsPrefix(prefix + "hdiv_residual_") # HDiv mass operator p = TrialFunction(V[self.vidx]) q = TestFunction(V[self.vidx]) mass = ufl.dot(p, q)*ufl.dx # TODO: Bcs? M = assemble(mass, bcs=None, form_compiler_parameters=self.cxt.fc_params) M.force_evaluation() Mmat = M.petscmat hdiv_mass_ksp.setOperators(Mmat) hdiv_mass_ksp.setUp() hdiv_mass_ksp.setFromOptions() self.hdiv_mass_ksp = hdiv_mass_ksp # Storing the result of A.inv * r, where A is the HDiv # mass matrix and r is the HDiv residual self._primal_r = Function(V[self.vidx]) tau = TestFunction(V_d[self.vidx]) self._assemble_broken_r = create_assembly_callable( ufl.dot(self._primal_r, tau)*ufl.dx, tensor=self.broken_residual.split()[self.vidx], form_compiler_parameters=self.cxt.fc_params) # Create the symbolic Schur-reduction: # Original mixed operator replaced with "broken" # arguments arg_map = {test: TestFunction(V_d), trial: TrialFunction(V_d)} Atilde = Tensor(replace(self.cxt.a, arg_map)) gammar = TestFunction(TraceSpace) n = ufl.FacetNormal(mesh) sigma = TrialFunctions(V_d)[self.vidx] # We zero out the contribution of the trace variables on the exterior # boundary. Extruded cells will have both horizontal and vertical # facets if mesh.cell_set._extruded: trace_bcs = [DirichletBC(TraceSpace, Constant(0.0), "on_boundary"), DirichletBC(TraceSpace, Constant(0.0), "bottom"), DirichletBC(TraceSpace, Constant(0.0), "top")] K = Tensor(gammar('+') * ufl.dot(sigma, n) * ufl.dS_h + gammar('+') * ufl.dot(sigma, n) * ufl.dS_v) else: trace_bcs = [DirichletBC(TraceSpace, Constant(0.0), "on_boundary")] K = Tensor(gammar('+') * ufl.dot(sigma, n) * ufl.dS) # If boundary conditions are contained in the ImplicitMatrixContext: if self.cxt.row_bcs: raise NotImplementedError("Strong BCs not currently handled. Try imposing them weakly.") # Assemble the Schur complement operator and right-hand side self.schur_rhs = Function(TraceSpace) self._assemble_Srhs = create_assembly_callable( K * Atilde.inv * self.broken_residual, tensor=self.schur_rhs, form_compiler_parameters=self.cxt.fc_params) schur_comp = K * Atilde.inv * K.T self.S = allocate_matrix(schur_comp, bcs=trace_bcs, form_compiler_parameters=self.cxt.fc_params) self._assemble_S = create_assembly_callable(schur_comp, tensor=self.S, bcs=trace_bcs, form_compiler_parameters=self.cxt.fc_params) self._assemble_S() self.S.force_evaluation() Smat = self.S.petscmat # Nullspace for the multiplier problem nullspace = create_schur_nullspace(P, -K * Atilde, V, V_d, TraceSpace, pc.comm) if nullspace: Smat.setNullSpace(nullspace) # Set up the KSP for the system of Lagrange multipliers trace_ksp = PETSc.KSP().create(comm=pc.comm) trace_ksp.setOptionsPrefix(prefix) trace_ksp.setOperators(Smat) trace_ksp.setUp() trace_ksp.setFromOptions() self.trace_ksp = trace_ksp split_mixed_op = dict(split_form(Atilde.form)) split_trace_op = dict(split_form(K.form)) # Generate reconstruction calls self._reconstruction_calls(split_mixed_op, split_trace_op) # NOTE: The projection stage *might* be replaced by a Fortin # operator. We may want to allow the user to specify if they # wish to use a Fortin operator over a projection, or vice-versa. # In a future add-on, we can add a switch which chooses either # the Fortin reconstruction or the usual KSP projection. # Set up the projection KSP hdiv_projection_ksp = PETSc.KSP().create(comm=pc.comm) hdiv_projection_ksp.setOptionsPrefix(prefix + 'hdiv_projection_') # Reuse the mass operator from the hdiv_mass_ksp hdiv_projection_ksp.setOperators(Mmat) # Construct the RHS for the projection stage self._projection_rhs = Function(V[self.vidx]) self._assemble_projection_rhs = create_assembly_callable( ufl.dot(self.broken_solution.split()[self.vidx], q)*ufl.dx, tensor=self._projection_rhs, form_compiler_parameters=self.cxt.fc_params) # Finalize ksp setup hdiv_projection_ksp.setUp() hdiv_projection_ksp.setFromOptions() self.hdiv_projection_ksp = hdiv_projection_ksp
import ufl from firedrake import * mesh = UnitSquareMesh(1, 1) Vhat = FunctionSpace(mesh, 'CG', 2) broken_elements = ufl.MixedElement( [ufl.BrokenElement(Vi.ufl_element()) for Vi in Vhat]) V = FunctionSpace(mesh, broken_elements) u = TrialFunction(V) v = TestFunction(V) mass = inner(u, v) * dx breakpoint() M = assemble(mass).M
def initialize(self, pc): """Set up the problem context. Take the original mixed problem and reformulate the problem as a hybridized mixed system. A KSP is created for the Lagrange multiplier system. """ from firedrake import (FunctionSpace, Function, Constant, TrialFunction, TrialFunctions, TestFunction, DirichletBC, assemble) from firedrake.assemble import (allocate_matrix, create_assembly_callable) from firedrake.formmanipulation import split_form from ufl.algorithms.replace import replace # Extract the problem context prefix = pc.getOptionsPrefix() + "hybridization_" _, P = pc.getOperators() self.cxt = P.getPythonContext() if not isinstance(self.cxt, ImplicitMatrixContext): raise ValueError("The python context must be an ImplicitMatrixContext") test, trial = self.cxt.a.arguments() V = test.function_space() mesh = V.mesh() if len(V) != 2: raise ValueError("Expecting two function spaces.") if all(Vi.ufl_element().value_shape() for Vi in V): raise ValueError("Expecting an H(div) x L2 pair of spaces.") # Automagically determine which spaces are vector and scalar for i, Vi in enumerate(V): if Vi.ufl_element().sobolev_space().name == "HDiv": self.vidx = i else: assert Vi.ufl_element().sobolev_space().name == "L2" self.pidx = i # Create the space of approximate traces. W = V[self.vidx] if W.ufl_element().family() == "Brezzi-Douglas-Marini": tdegree = W.ufl_element().degree() else: try: # If we have a tensor product element h_deg, v_deg = W.ufl_element().degree() tdegree = (h_deg - 1, v_deg - 1) except TypeError: tdegree = W.ufl_element().degree() - 1 TraceSpace = FunctionSpace(mesh, "HDiv Trace", tdegree) # Break the function spaces and define fully discontinuous spaces broken_elements = ufl.MixedElement([ufl.BrokenElement(Vi.ufl_element()) for Vi in V]) V_d = FunctionSpace(mesh, broken_elements) # Set up the functions for the original, hybridized # and schur complement systems self.broken_solution = Function(V_d) self.broken_residual = Function(V_d) self.trace_solution = Function(TraceSpace) self.unbroken_solution = Function(V) self.unbroken_residual = Function(V) # Set up the KSP for the hdiv residual projection hdiv_mass_ksp = PETSc.KSP().create(comm=pc.comm) hdiv_mass_ksp.setOptionsPrefix(prefix + "hdiv_residual_") # HDiv mass operator p = TrialFunction(V[self.vidx]) q = TestFunction(V[self.vidx]) mass = ufl.dot(p, q)*ufl.dx # TODO: Bcs? M = assemble(mass, bcs=None, form_compiler_parameters=self.cxt.fc_params) M.force_evaluation() Mmat = M.petscmat hdiv_mass_ksp.setOperators(Mmat) hdiv_mass_ksp.setUp() hdiv_mass_ksp.setFromOptions() self.hdiv_mass_ksp = hdiv_mass_ksp # Storing the result of A.inv * r, where A is the HDiv # mass matrix and r is the HDiv residual self._primal_r = Function(V[self.vidx]) tau = TestFunction(V_d[self.vidx]) self._assemble_broken_r = create_assembly_callable( ufl.dot(self._primal_r, tau)*ufl.dx, tensor=self.broken_residual.split()[self.vidx], form_compiler_parameters=self.cxt.fc_params) # Create the symbolic Schur-reduction: # Original mixed operator replaced with "broken" # arguments arg_map = {test: TestFunction(V_d), trial: TrialFunction(V_d)} Atilde = Tensor(replace(self.cxt.a, arg_map)) gammar = TestFunction(TraceSpace) n = ufl.FacetNormal(mesh) sigma = TrialFunctions(V_d)[self.vidx] if mesh.cell_set._extruded: Kform = (gammar('+') * ufl.dot(sigma, n) * ufl.dS_h + gammar('+') * ufl.dot(sigma, n) * ufl.dS_v) else: Kform = (gammar('+') * ufl.dot(sigma, n) * ufl.dS) # Here we deal with boundaries. If there are Neumann # conditions (which should be enforced strongly for # H(div)xL^2) then we need to add jump terms on the exterior # facets. If there are Dirichlet conditions (which should be # enforced weakly) then we need to zero out the trace # variables there as they are not active (otherwise the hybrid # problem is not well-posed). # If boundary conditions are contained in the ImplicitMatrixContext: if self.cxt.row_bcs: # Find all the subdomains with neumann BCS # These are Dirichlet BCs on the vidx space neumann_subdomains = set() for bc in self.cxt.row_bcs: if bc.function_space().index == self.pidx: raise NotImplementedError("Dirichlet conditions for scalar variable not supported. Use a weak bc") if bc.function_space().index != self.vidx: raise NotImplementedError("Dirichlet bc set on unsupported space.") # append the set of sub domains subdom = bc.sub_domain if isinstance(subdom, str): neumann_subdomains |= set([subdom]) else: neumann_subdomains |= set(as_tuple(subdom, int)) # separate out the top and bottom bcs extruded_neumann_subdomains = neumann_subdomains & {"top", "bottom"} neumann_subdomains = neumann_subdomains.difference(extruded_neumann_subdomains) integrand = gammar * ufl.dot(sigma, n) measures = [] trace_subdomains = [] if mesh.cell_set._extruded: ds = ufl.ds_v for subdomain in extruded_neumann_subdomains: measures.append({"top": ufl.ds_t, "bottom": ufl.ds_b}[subdomain]) trace_subdomains.extend(sorted({"top", "bottom"} - extruded_neumann_subdomains)) else: ds = ufl.ds if "on_boundary" in neumann_subdomains: measures.append(ds) else: measures.append(ds(tuple(neumann_subdomains))) dirichlet_subdomains = set(mesh.exterior_facets.unique_markers) - neumann_subdomains trace_subdomains.append(sorted(dirichlet_subdomains)) for measure in measures: Kform += integrand*measure trace_bcs = [DirichletBC(TraceSpace, Constant(0.0), subdomain) for subdomain in trace_subdomains] else: # No bcs were provided, we assume weak Dirichlet conditions. # We zero out the contribution of the trace variables on # the exterior boundary. Extruded cells will have both # horizontal and vertical facets trace_subdomains = ["on_boundary"] if mesh.cell_set._extruded: trace_subdomains.extend(["bottom", "top"]) trace_bcs = [DirichletBC(TraceSpace, Constant(0.0), subdomain) for subdomain in trace_subdomains] # Make a SLATE tensor from Kform K = Tensor(Kform) # Assemble the Schur complement operator and right-hand side self.schur_rhs = Function(TraceSpace) self._assemble_Srhs = create_assembly_callable( K * Atilde.inv * AssembledVector(self.broken_residual), tensor=self.schur_rhs, form_compiler_parameters=self.cxt.fc_params) schur_comp = K * Atilde.inv * K.T self.S = allocate_matrix(schur_comp, bcs=trace_bcs, form_compiler_parameters=self.cxt.fc_params) self._assemble_S = create_assembly_callable(schur_comp, tensor=self.S, bcs=trace_bcs, form_compiler_parameters=self.cxt.fc_params) self._assemble_S() self.S.force_evaluation() Smat = self.S.petscmat # Nullspace for the multiplier problem nullspace = create_schur_nullspace(P, -K * Atilde, V, V_d, TraceSpace, pc.comm) if nullspace: Smat.setNullSpace(nullspace) # Set up the KSP for the system of Lagrange multipliers trace_ksp = PETSc.KSP().create(comm=pc.comm) trace_ksp.setOptionsPrefix(prefix) trace_ksp.setOperators(Smat) trace_ksp.setUp() trace_ksp.setFromOptions() self.trace_ksp = trace_ksp split_mixed_op = dict(split_form(Atilde.form)) split_trace_op = dict(split_form(K.form)) # Generate reconstruction calls self._reconstruction_calls(split_mixed_op, split_trace_op) # NOTE: The projection stage *might* be replaced by a Fortin # operator. We may want to allow the user to specify if they # wish to use a Fortin operator over a projection, or vice-versa. # In a future add-on, we can add a switch which chooses either # the Fortin reconstruction or the usual KSP projection. # Set up the projection KSP hdiv_projection_ksp = PETSc.KSP().create(comm=pc.comm) hdiv_projection_ksp.setOptionsPrefix(prefix + 'hdiv_projection_') # Reuse the mass operator from the hdiv_mass_ksp hdiv_projection_ksp.setOperators(Mmat) # Construct the RHS for the projection stage self._projection_rhs = Function(V[self.vidx]) self._assemble_projection_rhs = create_assembly_callable( ufl.dot(self.broken_solution.split()[self.vidx], q)*ufl.dx, tensor=self._projection_rhs, form_compiler_parameters=self.cxt.fc_params) # Finalize ksp setup hdiv_projection_ksp.setUp() hdiv_projection_ksp.setFromOptions() self.hdiv_projection_ksp = hdiv_projection_ksp
from firedrake import * from firedrake.assemble import get_matrix, matrix_arg, get_vector, vector_arg from firedrake import tsfc_interface import numpy as np from numpy.linalg import cholesky import matplotlib.pyplot as plt # Set up RNG pcg = PCG64(seed=100) rg = RandomGenerator(pcg) # Mesh and FS mesh = UnitSquareMesh(2, 2) V = FunctionSpace(mesh, 'CG', 2) VB = FunctionSpace(mesh, ufl.BrokenElement(V.ufl_element())) samples = 1000 mass_size = (V.dof_count, V.dof_count) brok_size = (VB.dof_count, VB.dof_count) npmat = np.zeros(brok_size) Hnpmat = np.zeros(mass_size) for ii in range(samples): Wnoise = rg.normal(VB, 0.0, 1.0) HWnoise = Function(V) u = TrialFunction(V) v = TestFunction(V) # ~ ax[ii].matshow(Wnoise.dat.data.reshape(Wnoise.dat.data.size, 1)) #bcs = DirichletBC(V, 0, (1,2,3,4))
def white_noise(V, rng=None): ''' ''' # If no random number generator provided make a new one if rng is None: pcg = _default_pcg rng = RandomGenerator(pcg) # Create broken space for independent samples mesh = V.mesh() broken_elements = ufl.MixedElement([ufl.BrokenElement(Vi.ufl_element()) for Vi in V]) Vbrok = FunctionSpace(mesh, broken_elements) iid_normal = rng.normal(Vbrok, 0.0, 1.0) wnoise = Function(V) # We also need cell volumes for correction DG0 = FunctionSpace(mesh, 'DG', 0) vol = Function(DG0) vol.interpolate(CellVolume(mesh)) # Create mass expression, assemble and extract kernel u = TrialFunction(V) v = TestFunction(V) mass = inner(u,v)*dx mass_ker, *stuff = tsfc_interface.compile_form(mass, "mass", coffee=False) mass_code = loopy.generate_code_v2(mass_ker.kinfo.kernel.code).device_code() mass_code = mass_code.replace("void " + mass_ker.kinfo.kernel.name, "static void " + mass_ker.kinfo.kernel.name) # Add custom code for doing "Cholesky" decomp and applying to broken vector blocksize = mass_ker.kinfo.kernel.code.args[0].shape[0] cholesky_code = f""" extern void dpotrf_(char *UPLO, int *N, double *A, int *LDA, int *INFO); extern void dgemv_(char *TRANS, int *M, int *N, double *ALPHA, double *A, int *LDA, double *X, int *INCX, double *BETA, double *Y, int *INCY); {mass_code} void apply_cholesky(double *__restrict__ z, double *__restrict__ b, double const *__restrict__ coords, double const *__restrict__ volume) {{ char uplo[1]; int32_t N = {blocksize}, LDA = {blocksize}, INFO = 0; int32_t i=0, j=0; uplo[0] = 'u'; double H[{blocksize}*{blocksize}] = {{{{ 0.0 }}}}; char trans[1]; int32_t stride = 1; //double one = 1.0; double scale = 1.0/volume[0]; double zero = 0.0; {mass_ker.kinfo.kernel.name}(H, coords); uplo[0] = 'u'; dpotrf_(uplo, &N, H, &LDA, &INFO); for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) if (j>i) H[i*N + j] = 0.0; trans[0] = 'T'; dgemv_(trans, &N, &N, &scale, H, &LDA, z, &stride, &zero, b, &stride); }} """ # Get the BLAS and LAPACK compiler parameters to compile the kernel if COMM_WORLD.rank == 0: petsc_variables = get_petsc_variables() BLASLAPACK_LIB = petsc_variables.get("BLASLAPACK_LIB", "") BLASLAPACK_LIB = COMM_WORLD.bcast(BLASLAPACK_LIB, root=0) BLASLAPACK_INCLUDE = petsc_variables.get("BLASLAPACK_INCLUDE", "") BLASLAPACK_INCLUDE = COMM_WORLD.bcast(BLASLAPACK_INCLUDE, root=0) else: BLASLAPACK_LIB = COMM_WORLD.bcast(None, root=0) BLASLAPACK_INCLUDE = COMM_WORLD.bcast(None, root=0) cholesky_kernel = op2.Kernel(cholesky_code, "apply_cholesky", include_dirs=BLASLAPACK_INCLUDE.split(), ldargs=BLASLAPACK_LIB.split()) # Construct arguments for par loop def get_map(x): return x.cell_node_map() i, _ = mass_ker.indices z_arg = vector_arg(op2.READ, get_map, i, function=iid_normal, V=Vbrok) b_arg = vector_arg(op2.INC, get_map, i, function=wnoise, V=V) coords = mesh.coordinates volumes = vector_arg(op2.READ, get_map, i, function=vol, V=DG0) op2.par_loop(cholesky_kernel, mesh.cell_set, z_arg, b_arg, coords.dat(op2.READ, get_map(coords)), volumes) return wnoise
def white_noise(V, rng): ''' ''' # Create broken space for independent samples mesh = V.mesh() broken_elements = ufl.MixedElement( [ufl.BrokenElement(Vi.ufl_element()) for Vi in V]) Vbrok = FunctionSpace(mesh, broken_elements) iid_normal = rng.normal(Vbrok, 0.0, 1.0) wnoise = Function(V) # Create mass expression, assemble and extract kernel u = TrialFunction(V) v = TestFunction(V) mass = inner(u, v) * dx mass_ker, *stuff = tsfc_interface.compile_form(mass, "mass", coffee=False) mass_code = loopy.generate_code_v2( mass_ker.kinfo.kernel.code).device_code() mass_code = mass_code.replace("void " + mass_ker.kinfo.kernel.name, "static void " + mass_ker.kinfo.kernel.name) # Add custom code for doing "Cholesky" decomp and applying to broken vector blocksize = mass_ker.kinfo.kernel.code.args[0].shape[0] cholesky_code = f""" extern void dpotrf_(char *UPLO, int *N, double *A, int *LDA, int *INFO); extern void dgemv_(char *TRANS, int *M, int *N, double *ALPHA, double *A, int *LDA, double *X, int *INCX, double *BETA, double *Y, int *INCY); {mass_code} void apply_cholesky(double *__restrict__ z, double *__restrict__ b, double const *__restrict__ coords) {{ char uplo[1]; int32_t N = {blocksize}, LDA = {blocksize}, INFO = 0; int32_t i=0, j=0; uplo[0] = 'u'; double H[{blocksize}*{blocksize}] = {{{{ 0.0 }}}}; char trans[1]; int32_t stride = 1; double one = 1.0; double zero = 0.0; {mass_ker.kinfo.kernel.name}(H, coords); uplo[0] = 'u'; dpotrf_(uplo, &N, H, &LDA, &INFO); for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) if (j>i) H[i*N + j] = 0.0; trans[0] = 'T'; dgemv_(trans, &N, &N, &one, H, &LDA, z, &stride, &zero, b, &stride); }} """ cholesky_kernel = op2.Kernel(cholesky_code, "apply_cholesky", ldargs=["-llapack", "-lblas"]) # Construct arguments for par loop def get_map(x): return x.cell_node_map() i, _ = mass_ker.indices z_arg = vector_arg(op2.READ, get_map, i, function=iid_normal, V=Vbrok) b_arg = vector_arg(op2.INC, get_map, i, function=wnoise, V=V) coords = mesh.coordinates op2.par_loop(cholesky_kernel, mesh.cell_set, z_arg, b_arg, coords.dat(op2.READ, get_map(coords))) return wnoise